Characterizing the changing nature of African croplands

Keywords: agricultural change, remote sensing, deep learning, planet, Africa
Abstract Type: Paper Abstract

Abstract

African agriculture is undergoing widespread transformation in response to rapid economic and population growth, urbanization, and climate change. This transformation entails substantial cropland expansion, along with changing agricultural practices, such as growing numbers of medium-scale farms, increased cropping intensity, and other factors that are altering the continent’s smallholder-dominated agricultural systems. To understand the scale, pace, and extent of these changes, we developed annual, high resolution field boundary maps of croplands for 2018-2024 in three countries undergoing extensive agricultural change, Ghana, Zambia, and Tanzania. To develop the maps, we trained a U-Net, a convolutional neural network to map fields in high resolution Planet imagery, using >35,000 labelled images collected over 6 years across the continent. We used the trained model to predict the distribution of field interiors (as distinguished from fields edges and non-crop areas) for each year, and converted the resulting score maps into 1) estimates of cropping frequency over 6 years, and 2) maps of individual field instances. We validated the frequency maps using a pixel-based reference sample, along with a method to assess the consistency and reliability of annual field boundary maps. Using the validated maps to calculate zonal statistics within coarser grids (50, 100, 1000 m), we characterized agricultural systems in each of the three countries using 1) intensity, 2) temporal presence, 3) and median field size, to create a typology of croplands that distinguishes between small-, medium-, and large-scale systems, their cultivation intensity, and whether they occurring in zones of expansion, contraction, stasis, or evolution.